Unofficial code for the paper Certifying Some Distributional Robustness with Principled Adversarial Training
The code allows to train a ConvNet on MNIST using the adversarial training proposed in the paper.
Python 2.7, Tensorflow 1.3
model.py: class to build and train a ConvNet.
trainOps.py: class to train and test the model.
run_exp.sh: to launch an experiment.
exp_configuration: configuration file for an experiment, set here the hyperparameters.
Run
python download_and_process_mnist.py
to download and process MNIST dataset, then run
sh run_exp.sh
to start the training procedure, after having set the desired hyperparameters/configurations in exp_config and run_exp.